sapientinc/HRM ? reverse-engineered prompt

Build me the full project for this Hierarchical Reasoning Model in Python, so I can train and test it on puzzle style reasoning tasks like Sudoku, maze solving, and ARC. I want a working training script, evaluation script, dataset builders, and a simple way to visualize puzzle datasets in the browser. The main goal is to be able to generate or load the datasets, train the model from scratch, save checkpoints, and then evaluate accuracy on held out puzzles or a saved checkpoint. Please make it easy to run on one GPU or multiple GPUs, and include the config files and defaults needed to reproduce the example runs from the README. If anything depends on current library behavior, look up the latest docs online and use sensible defaults. Also make sure the code is organized cleanly enough that I can swap between the different puzzle datasets without rewriting everything.